<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Thet Twe Aung on Medium]]></title>
        <description><![CDATA[Stories by Thet Twe Aung on Medium]]></description>
        <link>https://medium.com/@thettwe?source=rss-7a425edbb2a9------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Esfd701V40IasNqwAbekqg.jpeg</url>
            <title>Stories by Thet Twe Aung on Medium</title>
            <link>https://medium.com/@thettwe?source=rss-7a425edbb2a9------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 14:24:36 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@thettwe/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[How I Converted 330K Burmese Words into 4.5 Million Romanized Spellings for a Pinyin-Style IME]]></title>
            <link>https://medium.com/@thettwe/how-i-converted-330k-burmese-words-into-4-5-million-romanized-spellings-for-a-pinyin-style-ime-dd418c648e85?source=rss-7a425edbb2a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/dd418c648e85</guid>
            <category><![CDATA[burglish]]></category>
            <category><![CDATA[nlp-training]]></category>
            <category><![CDATA[low-resource-language]]></category>
            <category><![CDATA[myanmar]]></category>
            <category><![CDATA[nlp]]></category>
            <dc:creator><![CDATA[Thet Twe Aung]]></dc:creator>
            <pubDate>Mon, 10 Aug 2026 03:16:58 GMT</pubDate>
            <atom:updated>2026-08-10T03:42:49.295Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-6JmBlFanzxeO8x5eUskWQ.png" /></figure><p>Burglish is the Burmese input method: the user types romanized Burmese by ear (nay kaung lar), and the engine returns ranked Burmese candidates (နေကောင်းလား). The asset that makes this possible is a large mapping from romanized keys to Burmese words, ranked by likelihood. No such mapping exists as public data. Romanized Burmese has no standard. The only grapheme-aligned romanization dataset I found contains 2,335 names and carries a non-commercial license: unusable on technical grounds and on licensing grounds alike.</p><p>I therefore generate the mapping myself, by converting every Burmese word in my lexicon into the romanized spellings a typist would plausibly produce. This report documents the source data, the conversion procedure, the techniques at each step, and the results.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-NyVs_L5qZGwH3t4Nkl4xg.png" /><figcaption><em>The pipeline: Burmese lexicon in, romanized key-word mapping out, compiled into one deterministic artifact the IME memory-maps read-only.</em></figcaption></figure><h3>Source data</h3><p>The primary source is the production database of a Myanmar spell checker (<a href="https://myspellchecker.com/">mySpellChecker</a>) I built previously. A spell checker and an IME share the same core task, mapping messy input to ranked candidates, so its data transfers directly. From its 18 tables I extract three: the word dictionary, the syllable inventory, and the word bigrams. Three external sources cover specific gaps. Each source is constrained so that what ships is extracted facts, counts and mappings, never any third-party source’s own text.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WlCjwYRRRbPXPzp5q8STqg.png" /><figcaption><em>The guards are load-bearing: four corpora contribute, and what crosses each boundary is derived data with clean provenance.</em></figcaption></figure><p>Data preparation applies four techniques:</p><ul><li><strong>Proof-based repair.</strong> Structurally impossible mark sequences are rejected (54 rows). The near-identical letters ဉ and ဥ are disambiguated by grammar: ဉ is a consonant, ဥ is an independent vowel, and only a consonant can carry a stacking mark, so ဥ under a stacking mark is provably a misspelling of ဉ and repairs mechanically. A legacy misencoding class that validates as Unicode but breaks canonical mark order (it had made the common word for “man” unreachable) is restored to canonical order before conversion.</li><li><strong>Cross-corpus frequency repair.</strong> 9,246 dictionary words carry frequency zero, meaning unknown, not unused. I repair 1,848 of them by borrowing counts from a permissively licensed segmented-corpus frequency list; borrowed counts are rescaled by the median frequency ratio, computed over the words both corpora attest.</li><li><strong>Weak labeling for loanwords.</strong> Wiktionary’s borrowing etymologies nominate candidate pairs, each an English key and a Burmese word. A pair ships only if the Burmese spelling is attested in my own lexicon, and it is my lexicon, not Wiktionary, that supplies the word and the frequency. The external source contributes one field, the English key, so the worst failure is a missing row, never a wrong one. The intersection cannot detect homograph false friends: king → ကင်း passes it, but ကင်း is a native word meaning &quot;devoid of,&quot; so the wrong pair would ship carrying a native word&#39;s frequency. A native-speaker pass over every harvested row rejected these.</li><li><strong>Zipf-based distillation.</strong> In my corpus the top 1,000 words carry 47.6% of usage, the top 10,000 carry 69.3%, and the top 150,000 carry 91%. I ship the top 300,000 words by frequency plus every curated word, 328,990 in all.</li></ul><p>A set-difference audit additionally proves that no shipped word is uniquely attributable to any non-commercially-licensed reference source: a word counts as at-risk only if it matches that source and appears in no cleanly licensed source.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qSZdOvLNlaRJYL8xeWixHQ.png" /><figcaption><em>Six stages, each with its row count; nothing is dropped without a recorded reason.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ns15sRQvi_jmKviu9Xg3Rg.png" /><figcaption><em>The cut point is an engineering decision: past it, added coverage stops paying for artifact size.</em></figcaption></figure><h3>The conversion procedure</h3><p>Burmese-to-romanization conversion cannot be learned, because no parallel data exists. It can be derived, because a Burmese syllable has a fixed, enumerable structure: an onset (a consonant, possibly aspirated, possibly carrying a medial glide), a rhyme (the vowel plus an optional final consonant), and a tone. Words are open-ended; the rhyme inventory is closed. In NLP terms the converter is a grapheme-to-phoneme system. Its output alphabet is Latin, and each sound is spelled the way attested Burglish variation spells that sound. The procedure per word:</p><ol><li>Segment the word into syllables.</li><li>Decompose each syllable into onset, medial, rhyme, and tone.</li><li>Romanize the parts compositionally against the rule tables.</li><li>Generate spelling variants.</li><li>Assign each generated key a confidence tier and emit key, tier, word, frequency rows.</li></ol><h3>Step 1: Syllable segmentation</h3><p>I implement canonical Myanmar mark order as a fifteen-slot state machine and segment against it. Three structural cases need explicit handling. The kinzi cluster is written on one syllable but sounds on the previous one, so when the segmenter reaches it, the converter goes back and rewrites a syllable it has already emitted. In a consonant stack, two consonants are written one above the other: the lower one is the pronounced onset of the next syllable, even though it sits under what reads as the current syllable’s coda. And NFC places the creaky-tone dot between a final consonant and its killer, the mark that cancels that consonant’s inherent vowel; a naive boundary rule would split the syllable between them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HW5pmJAIJymnRSk_BhdoNg.png" /><figcaption><em>ကောင်း occupies six of the fifteen slots; the validator rejects any syllable whose marks leave this order.</em></figcaption></figure><h3>Step 2: Decomposition and base romanization</h3><p>Three hand-authored rule tables drive romanization. An onset table (42 rows) maps each consonant to its by-ear roman. A rhyme table (144 rows) maps rhyme signatures to spellings; I define a rhyme signature as the codepoint sequence of vowel signs, coda, and killer with the two tone marks stripped. Rhyme rows list a primary spelling first, then attested alternates where the variation splits (လဲ appears as lai and lae). An equivalence table (43 substitution classes) records the swaps attested across spelling variants, such as hp ↔ ph ↔ p ↔ f. The classes include the aspirates of IAST, the standard scholarly transliteration for Sanskrit and Pali, which is how the scholarly spelling buddha reaches ဗုဒ္ဓ.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QdrfuK306z-ZnRVKnok2Aw.png" /><figcaption><em>Each syllable is sounded slot by slot, which is why the key matches what a typist hears rather than how the script spells it.</em></figcaption></figure><p>The phonological rules that produce by-ear keys rather than letterwise transliteration:</p><ul><li>The ha-medial marks aspiration, not an “h” sound.</li><li>အ is a silent zero onset: အိမ် keys as ein, never aein.</li><li>The wa-medial fuses with closed rhymes: ထွန်း keys as htun, with the letterwise htwan kept beside it.</li><li>Sandhi voicing: a stop that is not word-initial and sits between vowels is pronounced voiced, and the rule applies even through consonant stacks. So မန္တလေး yields mandalay, and မင်္ဂလာပါ becomes reachable from mingalarbar, the sandhi pronunciation.</li><li>Tone maps to typed conventions: bare ာ keys as a and the high-tone ား as ar (so စာ owns sa and စား owns sar), and the high tone is additionally typed as a doubled final letter (မေး also keys as mayy, at reduced confidence).</li></ul><p>The rule tables are asymmetric by design: some rows exist for recognition only, never for generation. Rare foreign codas carry a tag. The offline converter uses the tagged rows to romanize the existing loanwords that contain those codas, but the artifact compiler strips them from the packed rules it ships, so the runtime IME can read those spellings while it never learns to produce them.</p><p>Coverage is measured, not assumed: I run the converter over the full 20,300-syllable inventory, sort misses by frequency mass, fix the heaviest class, and repeat. Coverage weighted by how often each syllable occurs reached roughly 100% early. Coverage by distinct syllable types now stands at 91.5%, and the residual 8.5% of types is Mon and Pali codas whose aggregate frequency mass in the corpus rounds to zero.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YdWlPLzpp4JVolxx3ZkhJw.png" /><figcaption><em>Each climb is one batch of foreign-coda rhyme rows, added in frequency-mass order.</em></figcaption></figure><h3>Step 3: Variant generation</h3><p>Each word’s base key expands into a set of key families, groups of spelling variants that share one derivation. The tier a family lands in decides its rank weight later:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Xx_hpfolmHdCKt7So20eNA.png" /><figcaption><em>The complete key taxonomy per word. Tier 0 holds forms attested as typed; tiers 1 and 2 hold derived spellings at 4× and 16× frequency discounts.</em></figcaption></figure><p>he expansion runs under three controls:</p><ul><li><strong>Budgets by frequency rank.</strong> The top 100k words get 12 variants, the next 100k get 6, the tail gets 2. Introducing these budgets deleted 42% of all generated rows with zero movement on either evaluation set.</li><li><strong>Position-conditioned substitution.</strong> Equivalence classes apply at a word’s first and last matching position, and classes carry optional onset-only or final-only anchors. Typists write r for the consonant ရ, so an r↔y class is needed; but the r inside kar spells the rhyme ar, not a consonant, and a free r↔y class would rewrite it anyway. The positional anchor prevents that.</li><li><strong>Sandhi voicing with a veto and a threshold.</strong> Voiced variants are generated for non-initial onsets, and the fuzzy substitutions re-apply to the voiced forms. Two corrections bound this. First, a native-reviewed exclusion list of 697 key-word pairs removes voiced keys that belong to no reading: the same rules that correctly give တယ် the key dealso mint di for တည်, and which words actually voice is lexical, not derivable. Second, a voiced form may only displace the top of a shared key when its word is at least 2.0× more frequent than the incumbent. I froze the 2.0 threshold from a full-lexicon sweep in which genuine displacements sat at 4.2× or higher and near-ties sat around 1.8×.</li></ul><p>The converter’s actual output for one word shows the three mechanisms composing:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*nBSyiws5FGo6_AD4u5mdoA.png" /><figcaption><em>One faithful key, then sandhi voicing through the stack (</em><em>mandalay), tone and rhyme alternates, and the fuzzy substitutions re-applied to both forms.</em></figcaption></figure><h3>Step 4: Tiers and the ranking model</h3><p>I frame ranking as a noisy channel: the IME estimates P(word | typed) ∝ P(typed | word) · P(word). The conversion procedure plays the role of the channel model P(typed | word); since there is no parallel data to train on, its confidence is quantized into three tiers instead of being learned. P(word) is corpus unigram frequency, extended at runtime by a log-linear bigram factor with a floor for unseen pairs. The tier discount is a bit shift: stored weight = frequency &gt;&gt; 2·tier, a 4× discount per tier.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Rs6rgix1uoHvKCaCXZuLPA.png" /><figcaption><em>The scoring path. Quantizing the channel into tiers is what replaces training in a setting with no parallel data.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4pTF0EGYqktJVMU8WQk1eg.png" /><figcaption><em>Lower tiers trade rank for reach: a wilder spelling still finds its word, it just never outranks the faithful key.</em></figcaption></figure><p>Two constants in this model were set by measurement. The 4× discount was bracketed: 2× per tier lost on the tuning set, 8× lost on the held-out gate, 4× won on both. The discount exists because of a measured failure mode: a fuzzy variant of နဲ့ landed on the key nay and outranked နေ on raw frequency. Additionally, a within-key visibility floor prunes corpus echoes, entries that exist only because the corpus itself contains misspellings. Where the strongest word on a key is multi-syllable, any other entry on that key below 2% of that word&#39;s raw frequency is a misspelling of it, not a distinct reading, and is dropped. Single-syllable keys are never pruned: offering alternative readings there is the IME&#39;s function.</p><h3>Step 5: The curated layer</h3><p>Knowledge that rules cannot derive ships as data. The complete rule layer, meaning the compositional tables, the generated loanword keys, and the corrections combined, comes to 1,816 reviewable rows:</p><ul><li><strong>Curated typed forms</strong> with no compositional path, such as the contracted pronoun ကျွန်တော် typed kyunto, unreachable from the orthographic kywantaw by any single substitution.</li><li><strong>56 standalone-intent pins</strong> for keys where corpus frequency favors a bound morpheme (on ka, the subject marker က outweighs the noun ကား; the pin restores the noun for the standalone case). Pins stay scoped to standalone input: a tested change that strengthened pins inside phrase composition raised the curated gate score but dropped real-sentence accuracy from 21.9% to 14.0%, and was reverted.</li><li><strong>A 21-word sentence-final particle class</strong> whose bigrams receive a boost, correcting a bias in the bigram source: it is news text, and in news the continuation ကြီး·လာ outcounts the chat question ကြီး·လား.</li><li><strong>The exclusion lists</strong> from Step 3, plus the adjudicated loanword false friends.</li></ul><p>The authoring method for every table is the same: apply the general rule, measure what it breaks, enumerate the affected class, and adjudicate with a native speaker. One measured example: a proposed convention for a homophone series produced 101 colliding keys on enumeration, including the keys for ဖုန်း (273k frequency) and သုံး (385k), both of which the rule would have broken. Five adjudicated pins shipped instead; a softer class-wide variant was also tested, regressed both evaluation sets, and was reverted.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ytVdzPY12eGyWqsEbgPMag.png" /><figcaption><em>The largest table is the veto list: correcting over-generation costs more rows than generating.</em></figcaption></figure><h3>Compilation</h3><p>The generated mapping is 4,568,088 key-tier-word-frequency rows. It compiles offline into one 43.6 MB artifact with five sections: packed rule tables, a lexicon finite-state transducer (FST) with posting lists, the bigram table, the English suppression set, and metadata carrying the data version. The IME loads the artifact fail-closed against a checksum:</p><ul><li>The FST maps each key to a posting list capped at 16 words, each with a one-byte weight relative to the list’s strongest entry (255 = strongest). Identical posting lists deduplicate to one stored copy; most fuzzy variants of a word share their list, and this deduplication took the first compile from 90.4 MB to 19.9 MB.</li><li>The bigram table stores no text: 250,000 hashed pairs, nine bytes per pair, counts log-quantized to one byte, both words canonicalized before hashing so counts split across spellings combine.</li><li>Every word passes a canonical-mark-order validator. Most Myanmar combining marks carry combining class zero, and NFC never reorders class-zero marks; a wrong order is therefore still valid Unicode, and it renders as garbage. Validation is the only reliable gate.</li><li>The build is byte-deterministic. The data version embeds a fingerprint over every rule table, and all ranking constants live in one manifest read by both the Python converter and the Rust compiler, so neither a rule edit nor a drifted constant ships unnoticed. After encoding, the compiler streams every key back out of the finished artifact and re-validates every decoded word, a check the memory-mapped runtime cannot afford at load.</li><li>One fixture file drives every language binding: 17 conversion cases with pinned expectations plus 7 synthetic result shapes, run identically through the C ABI, through JNI inside an embedded JVM, and through the WebAssembly build under Node. CI adds a fourth consumer: a real C program, compiled on every run against a header regenerated from the Rust source. Five of the synthetic shapes are deliberately malformed, among them a misordered ranking and a syllable out of canonical order, and the suite fails if the checker accepts any of them, so the checker is under test too.</li></ul><h3>How the IME consumes it</h3><p>The offline build is a procedure, which is why I described it as steps. The runtime is better described as three judgments the engine makes on every keystroke over the memory-mapped artifact: finding the candidates, refusing the impostors, and ordering the window. The last part is the bill for making them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p9TWPojQW5GniFoAS_rd3g.png" /><figcaption><em>The path per keystroke. The by-ear expansion and the lattice cover reuse the same rule tables and bigram model the offline build packed.</em></figcaption></figure><h3>Finding the candidates</h3><p>The typed buffer is normalized first: chat elongation such as kaungggg collapses, while doubled letters survive, since doubling carries tone. The key then expands into a bounded set of lookups: exact, despaced, and by-ear variants from re-applying the equivalence classes up to two substitutions deep, with each depth banded strictly below the previous one, so a two-edit match can never outrank a direct hit. Lookup itself is a single automaton walk per lattice position, and syllable spans are capped at 32 bytes against a true maximum near 11, so lattice construction stays linear in buffer length and a pasted URL cannot make a keystroke quadratic. A whole-buffer dictionary hit returns its ranked posting list directly. A buffer that misses is instead covered by a lattice segmenter: it splits the buffer into the fewest dictionary words that cover it, scoring the joints with the bigram model. That segmenter path is how naykaunglar resolves into three words.</p><h3>Refusing the impostors</h3><p>Recall this wide needs a discipline for saying no. A stitched cover, a candidate the segmenter assembled from several dictionary words, is offered only with evidence the words belong together: either the token segments cleanly as romanized Burmese syllables, or every adjacent pair of the stitched words is attested in the bigram table. Every pair, not just one: letter salad over a foreign word can ride a single coincidentally real pair. I also tried waiving the check for run-ons that begin with a foreign word, and rejected the exemption because it readmitted the same garbage for English names. Synthesis, the path that builds a candidate from the rule tables rather than the dictionary, is bounded the same way: it produces sounds only, never written structure the romanization cannot encode, so kinzi and the great-sa ligature come from the dictionary alone.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*R1mwN8aT2--7aEje6i5ZnQ.png" /><figcaption><em>Refusal is a feature: an IME that converts everything would bury honest candidates under confident nonsense.</em></figcaption></figure><h3>Ordering the window</h3><p>Candidates rank in strict provenance bands, exact above by-ear above stitched above synthesized, with the literal Latin echo reserved in the last slot as the safety net. Two signals then adjust the cold order, and each carries a guard.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8Phzr0vr7rgV_J5XFFO12g.png" /><figcaption><em>Both adjustment signals carry guards: pins are exempt from context, and forced commits teach nothing.</em></figcaption></figure><p>Context is contrastive, not a raw bigram multiply. Each candidate gets a bigram strength with the previously committed word; that strength is then compared to the mean strength across the whole candidate set, and the candidate’s score is scaled by 0.03^((mean − strength)/255). The effect: a reading the previous word predicts above average is lifted past its cold score, not merely allowed to lose less than the others. Against the demotion-only version, which can push weak candidates down but never lift a strong one, the contrastive form gained 2.3 points on the context corpus and 13.9 points on left-separable homographs, the homographs whose reading the preceding word decides. The demotion-only version stays exposed so the accuracy rig can A/B the two.</p><p>Habit is a personal bigram model over the user’s own commits, with Katz-style backoff between per-context and global distributions. A per-context bucket takes over from the global one only once a single reading holds an 80 percent share of at least 1.5 decayed picks in that context; that threshold is how the same typed key settles on ချင် after လို and ခြင်း after နိုင်. A stray pick cannot do what a habit can, for three reasons: the learned bonus follows a bounded curve in the decayed pick count, so one pick buys little; a finalize forced by focus loss records at weight zero, so it teaches nothing; and every learned strength halves after each 500 commits, counted on the store’s own commit counter rather than wall-clock time.</p><h3>What it costs, measured</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ulxnu9lge5jpYBiHNDh9rg.png" /><figcaption><em>Nothing here depends on a server behaving well: every guarantee is local and measured on the shipped build.</em></figcaption></figure><h3>Evaluation and results</h3><p>The accuracy gate is a frozen set of 266 human-typed inputs with adjudicated expected outputs, held disjoint from everything that feeds the lexicon and the rules, enforced as a no-regression CI check. I froze it only after three separate models had independently reviewed every row it held at that point. Their pairwise agreement ranged from 98.8 to 99.6%, and no row was flagged by more than one of the three. Contamination is treated as the primary threat, and two mechanisms guard against it. First, the harness reports a compositional score that excludes every row whose answer a pin or curated entry forces, so hand-curated fixes cannot inflate the derived-rule number. Second, the CI number runs on a committed sample lexicon, the closure of everything the tests reference, rather than the full lexicon; that difference is why sample and full-lexicon scores are reported separately. Saturation, a gate scoring near its ceiling and no longer discriminating, is a real and measured failure mode: an earlier 102-row version of the gate read 99% and dropped 21 points when independently authored rows were added.</p><p>The deployment metric is per-token recall@9 on real chat sentences: each typed token is aligned to its expected syllable span by dynamic programming, and the question is whether the intended word appears in the 9-candidate window.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*REqHmwZX-Y-0TvWUZkgacA.png" /><figcaption><em>One engine, five honest measurements: which number to quote depends on which promise you are checking.</em></figcaption></figure><p>Each gate row is tagged with the linguistic phenomena it exercises, so the score can be broken down to show where the converter is strong and where it is weak. Selected full-lexicon top-1 by tag (a row can carry several tags):</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LxZDZFx4g7m3D6WxnCk6pA.png" /><figcaption><em>Per-tag scores turn one average into a diagnosis: they show which phenomena the rules have mastered and where the remaining misses cluster.</em></figcaption></figure><p>Output totals: 326,617 of 328,990 distilled words romanized into 4,568,088 key-word rows; 485,571 unique primary keys, of which 14,170 map to more than one word. That ambiguity is the intended product shape: nay reaches နေ, နေ့, နေး, and kyaung carries ကျောင်း and ကြောင် with their tones, ranked by discounted frequency. One key serving several ranked words is exactly what lets a typist use any plausible spelling without learning a convention.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vSVVe2pzlGnUu1bSpH1eNg.png" /><figcaption><em>The discount is what lets 3.6 million guesses coexist with half a million attested forms.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*PdbRpRmkW6d9qmBri3vHig.png" /><figcaption><em>kyaung alone carries the school, the cat, and their tones.</em></figcaption></figure><p>The generated mapping now exists as a reproducible build: one deterministic 43.6 MB artifact, rebuilt bit-identically from the manifest-pinned sources, carrying the romanization layer that written Burmese never had.</p><p>Everything measured here ships inside the keyboard itself: Burglish is free to download at <a href="https://burglish.app/">burglish.app</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dd418c648e85" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Hell yeah! .. Gawdai is finally on Medium!]]></title>
            <link>https://medium.com/gawdai/hell-yeah-gawdai-is-finally-on-medium-4925c5e2587d?source=rss-7a425edbb2a9------2</link>
            <guid isPermaLink="false">https://medium.com/p/4925c5e2587d</guid>
            <category><![CDATA[myanmar]]></category>
            <category><![CDATA[apps]]></category>
            <category><![CDATA[updates]]></category>
            <category><![CDATA[announcements]]></category>
            <category><![CDATA[gawdai]]></category>
            <dc:creator><![CDATA[Thet Twe Aung]]></dc:creator>
            <pubDate>Wed, 21 Sep 2016 13:34:47 GMT</pubDate>
            <atom:updated>2016-09-22T02:07:29.894Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FkIXySfsKQ206As772x9fw.jpeg" /></figure><p>Well we know that we shouldn’t be so much hype about this. But with tiny little schedules that we have, it’s a huge commitment for us to plan for writing. So let us hype for awhile until we cool down. Yipeeeee! 🎉</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fgiphy.com%2Fembed%2FwAxlCmeX1ri1y%2Ftwitter%2Fiframe&amp;display_name=Giphy&amp;url=https%3A%2F%2Fgiphy.com%2Fgifs%2Fromy-dance-dancing-snoop-dogg-wAxlCmeX1ri1y&amp;image=https%3A%2F%2Fmedia1.giphy.com%2Fmedia%2FwAxlCmeX1ri1y%2Fgiphy.gif&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=giphy" width="435" height="326" frameborder="0" scrolling="no"><a href="https://medium.com/media/07b03fd6ef7459696c4f75e45454e5fa/href">https://medium.com/media/07b03fd6ef7459696c4f75e45454e5fa/href</a></iframe><p>We have so many reasons to start writing. Among them, a sole reason that keeps pushing us is to start sharing detailed insights of what’s happening in Gawdai.</p><p>Here are the key topics that you could expect reading regularly on our blog.</p><ol><li>Product Roadmap</li><li>Features Highlights</li><li>Articles about Linguistics &amp; Slang</li><li>Behind the scene works</li><li>Tutorials</li><li>and a lot of other things that would worth your time for a read</li></ol><p>Numbers of the <strong>announcement</strong> are up our sleeves to reveal. But in this post, we will go with the features update first.</p><p>Yes, Gawdai was very silent in past few months. That was because we had been working very hard to bring the <strong>version 2</strong> to you.</p><p>In version 2, we focus on stability as well as the improvements &amp; enhancements. Let us present some of the new features that you are going to enjoy in this Version 2.</p><ol><li><strong>Search</strong></li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*0BM3vJuRSs-3bN58G_MLfg.gif" /><figcaption>Search Features on Gawdai Web</figcaption></figure><p>We know right, this feature should have been included from the day 1. We are fully aware that a lot of our user has been longing for it and we want this so bad as much as the user do. The reason we couldn’t bring search capabilities in our 1st release is due to the tight of our implementation time as well as the difficulties to facilitate both Zawgyi &amp; Unicode font (yes, the struggle is real, we will go details about how we successfully serve both Unicode and Zawgyi users in our future post together with our commitments &amp; struggles, keep an eye on that).</p><p>So yeah, finally we made it. Search features will be available on all the platforms (Web, iOS &amp; Android). Web user can start enjoying the perks of this feature and it will come to mobile users very soon. The user can search by any keywords and also locate the slang words directly through auto-suggestion.</p><p><strong>2. Browse by Alphabets</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/750/1*QmAyvbo7svPIuFAl9kjehA.jpeg" /><figcaption>Browse by Alphabet Features on Gawdai iOS App</figcaption></figure><p>We always wonder if we could browse slangs by alphabets (inspired by <a href="http://urbandictionary.com/">UrbanDictionary</a>) that would be really cool. So instead of idling in wonder land, we get up and bring this out to our user. This might not be the “wow” feature for some, but this is a very big leap for the usability as well as the content’s discoverability. Enjoy browsing!</p><p><strong>3. Happenings</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*udamcKqi388HiDjwYVo7Zw.png" /><figcaption>Happening Features on Gawdai Web</figcaption></figure><p>Gawdai main interest has always been in the community and linguistic evolution. The community is a life source of the Gawdai and we really appreciate of how everyone shaping up Gawdai to become what it is today. With the thoughts of how we could bring the community of Gawdai closer, we bring this “Happenings” features as the first step. Currently, it only focuses on the capturing of user’s activities to bring more user’s engagement and content’s discoverability. In future, we will introduce social interaction features such as “Follow” and “Notification” of the user who you followed.</p><p><strong>4. Voting Improvements</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/764/1*Ll3AkIAkeKZgQXxWCiU9_g.gif" /><figcaption>Voting Improfements on Gawdai Web</figcaption></figure><p>Voting plays a major role in conducting the quality of the content that user submitted. Previously, we allowed public voting but some malicious users abused the voting process to break the balance of the community. Because of that, we are decided to set the voting process available for the registered/logged-in user only. Along with it, we have worked on some of the improvements on the voting process as well. Now, user can undo the accidental actions. Neat right?</p><p><strong>5. Kill em all</strong></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fgiphy.com%2Fembed%2FwvFrJa4aWKuvm%2Ftwitter%2Fiframe&amp;src_secure=1&amp;url=http%3A%2F%2Fgiphy.com%2Fgifs%2Ffire-with-kill-wvFrJa4aWKuvm&amp;image=https%3A%2F%2Fmedia.giphy.com%2Fmedia%2FwvFrJa4aWKuvm%2Fgiphy.gif&amp;key=d04bfffea46d4aeda930ec88cc64b87c&amp;type=text%2Fhtml&amp;schema=giphy" width="435" height="326" frameborder="0" scrolling="no"><a href="https://medium.com/media/89131719f34ec2d42dd8b27179291753/href">https://medium.com/media/89131719f34ec2d42dd8b27179291753/href</a></iframe><p>System/Software bugs crawl like other bugs do. Before we know it, they will be all over the place. We constantly monitor and kill them whenever we got a chance to since stability is the user’s satisfaction. In this version, we’ve fixed numbers of spotted bugs and user will experience the more mature and stable version that ever produced by us.</p><p>These are just the overview highlights and we didn’t stop it there. As we mentioned above, we have quite a number of <strong>big announcements</strong> to make and we don’t want you to miss that. For now, just sit back, relax and enjoy the new version of Gawdai.</p><blockquote>Keep Gawing guys. It’s just the beginning. ✊🏼</blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4925c5e2587d" width="1" height="1" alt=""><hr><p><a href="https://medium.com/gawdai/hell-yeah-gawdai-is-finally-on-medium-4925c5e2587d">Hell yeah! .. Gawdai is finally on Medium!</a> was originally published in <a href="https://medium.com/gawdai">Gawdai</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>